For the first time, an AI system has autonomously discovered a genuinely new mathematical result that human experts find significant in itself. In May 2026, OpenAI reported that a reasoning model produced the core ideas to disprove the Erdős unit-distance conjecture, an unsolved problem in discrete geometry that had stumped mathematicians since 1946. The AI found point configurations with far more unit-distance pairs than the conjecture allowed, showing the count can grow at least as fast as n^1.014. The result was verified by external mathematicians, including Fields Medalist Tim Gowers.

"The first example of a result produced autonomously by an AI that I find exciting in itself," said Tim Gowers.

Tim Gowers, Fields Medalist and Mathematician

This breakthrough marks a watershed moment in AI research. For years, the honest answer to "Can AI create genuinely new knowledge?" was "not really; it remixes what it's seen." But in mid-2026, that changed. The practical implication is profound: research-grade reasoning is now a capability you can access through an API, useful for hard optimization problems, novel algorithm design, and questions with no known answer to memorize.

What Does AI Discovery Actually Mean for Math and Science?

The conjecture disproof is not an isolated incident. Google DeepMind released AI Co-Mathematician in May 2026, an interactive workbench of AI agents that supports the full research workflow, including ideation, literature search, computational exploration, and theorem proving. The system maintains state across a session, tracking failed hypotheses instead of starting fresh with each prompt. It scored 48% on FrontierMath Tier 4, the hardest tier of the benchmark, a new high among systems tested at the time.

Even more striking, a two-agent system called the Automated Conjecture Resolution framework resolved an open problem in commutative algebra with essentially no human involvement. One agent searches for proofs while a second agent formalizes them in Lean 4, a formal verification language. This means every result is machine-checked line by line. The team found new counterexamples in algebraic groups and p-adic Hodge theory, all formally verified so the AI doesn't ask you to trust its proof; it produces one a computer can verify.

How Are AI Agents Becoming More Reliable for Real Work?

While AI discovery captures headlines, quieter shifts matter more for anyone building software. Researchers discovered that agents got measurably better at choosing the right tool, and the cost of running large models dropped by up to 100x in specific settings. A 30-billion-parameter open model reached 64% on a real coding benchmark, showing that smaller models can punch above their weight class.

Tool selection has been a persistent weak point for AI agents. A training-free method called Tool-DC ("Try, Check, and Retry") uses a divide-and-conquer loop that lets a model iteratively narrow down the right tool from a large candidate set using self-reflection. The training-free version delivered up to 25.10% average gains on tool-calling benchmarks. The training-based variant pushed a Qwen2.5-7B model to 83.16% on BFCL, past OpenAI o3 and Claude Haiku 4.5 on that test. For builders with dozens or hundreds of tools, this means you don't need a fine-tuned router; a plug-in decompose-and-verify wrapper can cut tool-selection errors and lift a small open model to proprietary-tier function-calling accuracy.

Another critical finding: agents still can't reliably debug themselves. A benchmark called AgentHallu tested whether models can pinpoint which step in a multi-step run caused a hallucination and why. Across 13 leading models, the best reached only 41.1% step-localization accuracy. Tool-use hallucinations were the hardest to catch, at just 11.6%. This is a reality check for builders: don't rely on the agent to attribute its own failures. You still need explicit step-level tracing and evaluation harnesses.

Steps to Build More Reliable AI Agents

• Implement External Verification Layers: Pair a generator with a checker in a "GAN-style" loop, as demonstrated by the Automated Conjecture Resolution framework. This pattern generalizes far beyond proofs to code, data pipelines, and any agent output that needs to be trusted.
• Use Decompose-and-Verify Wrappers for Tool Selection: Instead of fine-tuning a router, apply Tool-DC's training-free divide-and-conquer approach to narrow down the right tool from large candidate sets, improving accuracy by up to 25% without additional training.
• Add Explicit Step-Level Tracing and Evaluation: Since models are poor at self-diagnosing where agents went wrong, especially at tool-use steps, build in external monitoring and adversarial evaluation harnesses rather than relying on the agent's own self-checks.
• Run Concurrent Reactive and Planning Tracks: For live environments like trading, ops monitoring, or real-time support, use AgileThinker's dual-track architecture to run a fast reactive head and a deeper planner simultaneously, rather than forcing a choice between latency and depth.

Why Is Cost Suddenly Dropping So Dramatically?

The most underrated research thread of 2026 isn't capability; it's cost. Three results point the same direction: stop throwing raw compute at the problem. TurboQuant, a training-free method from Google Research and DeepMind, compresses the LLM key-value cache to about 3.5 bits per channel with no measurable quality loss, using random rotation plus optimal scalar quantization. This achieves roughly a 6x reduction in KV-cache memory without sacrificing performance.

These cost reductions matter because they democratize access to frontier-grade reasoning. If you can run a model 6x more efficiently, or reduce inference costs by 100x in specific settings, the economics of AI applications shift dramatically. Smaller teams can now afford to deploy models that were previously accessible only to well-funded labs.

The convergence of these findings tells a story: AI research in mid-2026 is moving from "Can we build this?" to "Can we build this reliably and affordably?" The discovery of new mathematics is the headline, but the real transformation is in the infrastructure, cost, and reliability of AI systems that builders can actually use.